Alright, so I ended up running GPUGRID on an AMD EPYC 7401P thanks to Packet’s (Packet.net) trial they gave me.

I ran this “cloud server” for around 5-6 days and loaded GPUGRID and Rosetta to see how much crunching power you can get for 1 [dollar/hour] (normal cost for this 24-core setup). I’ve already had the pleasure of playing with Google Compute, but I think their servers are “virtual”, unlike Packet’s solution that is a literal 24-Core CPU all for yourself. Side note: It would be interesting to see what you could get for 1 [dollar/hour] 5 years in the future!

Anyways, since I ran two projects at the same time, I had to do some data tinkering to separate the EPYC’s performance on Rosetta and GPUGRID. What I did was download the PC’s tasks results into a spreadsheet (I actually had to use VBA, copy-paste 20 results at a time… there is no “download” button on the results page!) and did some good ol' maths.

I got the following results:

Take note that this CPU has 48 threads in total. Remember I ran this alongside Rosetta.

Measured floating point speed: 3,574.13 million ops/sec
Measured integer speed: 76,379.83 million ops/sec
Total Credits (50/50 Rosetta): 177,901
Total CPU Time: 10,424,024 seconds (about 120 days)
Run time: 2,735,510 seconds (about 32 days)… since each WU uses 4 threads, the CPU time is 4 times this number.
“Real” Run time: 383,373 seconds (time between first WU sent and last WU reported)
Credits per Second per Thread: 0.0171
Theoretical Credits per Second per WU: 0.065
Theoretical Credits per Second: 0.78 (0.065 x 12 since this CPU can run 12 WUs in parallel)
Credits per Day according to my experiment (50/50 Rosetta): 40,093
Theoretical Credits per Day (if running 100% GPUGRID): 70,232

I graphed the CPU time versus the awarded credits (blue) and Credits/second (red). This is using “real” CPU time (ignoring the fact that each WU uses 4 cores). Seems the amounts of credits awarded increases linearly with time.



I wouldn’t know how much faster/productive this CPU is compared to others on this project, but it seems to be a beast. About 70k per day.


Anyone care to calculate their Credits/second on their CPU to compare?

---

~30000 credit on my R7 @ 3.2GHz (the only system in my profile)
0.021812 credit per second per thread

I wouldn’t know how much faster/productive this CPU is compared to others on this project, but it seems to be a beast. About 70k per day.

Amd makes a great job with these cpu.
And during 2019 will be released the "Rome" cpu-server and Ryzen 2 for cpu-client.
Competition is welcome!